Hugging Face bans hacker version of GLM-5.3, author reuploads after renaming

Beating AI News Flash: Hugging Face has banned penclaw-GLM-5.3-abliterated-for-offensive-cyber, uploaded by AI security company Audn, with the page showing it violated the Content Policy. This version directly modifies the GLM-5.3 weights, weakening the model's refusal behavior. The original repository name and description also explicitly stated offensive cyber. The author subsequently deleted those words and re-uploaded the model.

Hugging Face has not publicly explained exactly why it was banned. The author claims they also do not understand the reason, and some in the community speculate that the offensive cyber naming triggered the review. Hugging Face's current policy indeed prohibits content intended to damage, gain unauthorized access to systems, and generate malicious code.

But versions that weaken model refusal restrictions are not uncommon on Hugging Face, and there are already thousands of similar abliterated models on the platform.